The DIVA (Directions Into Velocities of Articulators) model is an adaptive neural network model that is used to control the movement of the analog vocal tract to generate words, syllables, or phonemes. The input signal to the DIVA model is the EEG (electroencephalogram) signal acquired from the human brain. However, due to the influence of power frequency interference and other forms of noise, the input signal can be non-stationary and can also contain a variety of multi-form waveforms in its instantaneous structure. Input of such a signal into the DIVA model affects normal speech processing. Therefore, based on the concept of sparse decomposition, this paper applies and improves an adaptive sparse decomposition model for feature extraction of the general EEG signal structure and then uses the Matching Pursuit algorithm to compute the optimal atom. The original EEG signal can then be represented by atoms in a complete atomic library. This model removes noise from the EEG signal resulting in a better signal than the wavelet transform method. Finally, applies the EEG signal de-noised by this model to DIAV model. Simulation results show that the method improves phonetic pronunciation greatly.